Abstract: Surface wave suppression directly influences the imaging quality and interpretation accuracy of seismic data. In view of the limited identification accuracy and generalization of intelligent noise suppression based on a completely data-driven deep learning architecture, a physically constrained surface wave suppression technique is proposed. To enhance noise recognition, a UNET architecture with dual-channel input, jointly from the time-space domain and corresponding frequency-wavenumber domain before denoising, is constructed to establish the mapping relationship between input data and output noise data in the time-space domain. Based on the characteristics of surface waves as regular noises, structural similarity regularization operators are introduced into the loss function to further enhance the network's ability to recognize noises. To address different physical characteristics of surface waves in different work areas, noise frequency and apparent velocity distributions are used as the constraints for further processing of output noise data to obtain higher-precision noise predictions, which will be subtracted from input data using an adaptive subtraction algorithm to obtain final denoised data. The testing on several field data sets shows superior denoising accuracy and generalization capability of the proposed algorithm.